DGMiniNet:Dynamic Graph Convolution Network for LiDAR Point Cloud Semantic Segmentation on Spherical coordinates

Abstract

Semantic segmentation of LiDAR data is a vital activity aiming at giving a semantic category to each point in the LiDAR point cloud, allowing accurate recognition and segmentation of objects and structures in the environment.We can obtain detailed information about semantic categories like roads, buildings, vehicles, pedestrians, etc. by classifying the points in the point cloud. This information can be used to support decision-making in the areas of autonomous driving, intelligent transportation systems, environment sensing, and scene understanding.In this research, we propose DGMiniNet, a spherical projection-based approach for semantic segmentation of LiDAR. First, a 2D representation is learned after spherically projecting the 3D LiDAR point cloud.Then, DGMiniNet receives the 2D representation.Creating feature maps and using 2D fully convolutional neural networks (FCNN) to generate semantic labels in 2D space.The labels that have been given in two-dimensional space are then reprojected into three-dimensional space via the post-processing module.Accurate point cloud contour extraction is accomplished by enhancing the association of dynamic feature maps with labels while simultaneously adding information from semantic segmentation, resulting in more semantically intelligible and geometrically accurate outputs.The semantickitti dataset was used to train our method, which has a 93% accuracy rate.